def call(self, *args: tf.Tensor, **kwargs: tf.Tensor) -> tf.Tensor:
"""Convert input tensors arguments into a signal tensor."""
# Don't use `training` or `mask` arguments from keras.Layer.
for k in ['training', 'mask']:
if k in kwargs:
_ = kwargs.pop(k)
controls = self.get_controls(*args, **kwargs)
signal = self.get_signal(**controls)
return signal
def parse_example(self, example: tf.Tensor):
example = super().parse_example(example)
image = example.pop("image")
if self._aug:
image = tf.image.random_flip_left_right(image)
image = tf.image.random_flip_up_down(image)
image = random_rot(image)
image = tf.image.random_brightness(image, max_delta=0.5)
image = tf.image.random_contrast(image, 0.5, 1.5)
image = tf.image.random_hue(image, max_delta=0.5)
image = tf.image.random_saturation(image, 0.5, 1.5)
image = random_resize_crop(image)
image = tf.image.resize(
image, (224, 224),
preserve_aspect_ratio=False,
antialias=False
)
label = example.pop("label")
return image, label
def parse_example(self, example: tf.Tensor) -> Dict[str, tf.Tensor]:
example = super().parse_example(example)
example.pop("image_id")
example.pop("image_type")
image = example.pop("image_content")
seg = example.pop("segment_mask")
if self._shp is not None:
seg = seg[..., tf.newaxis]
seg = tf.image.resize(
seg,
size=self._shp,
method=tf.image.ResizeMethod.NEAREST_NEIGHBOR,
antialias=False)
seg = seg[..., 0]
# one-hot classes
if self._one_hot_class:
seg = tf.one_hot(seg, depth=self.nClass)
else:
seg = seg[..., tf.newaxis]
example["image"] = image
example["segment_mask"] = seg
return example
def parse_example(self, example: tf.Tensor) -> Dict[str, tf.Tensor]:
example = super().parse_example(example)
example.pop("image_id")
example.pop("image_type")
image = example.pop("image_content")
label = example.pop("classes")
example["image"] = image
example["label"] = label
return example
def parse_example(self, example: tf.Tensor) -> Dict[str, tf.Tensor]:
example = super().parse_example(example)
example.pop("image_id")
example.pop("image_type")
image = example.pop("image_content")
boxes = example.pop("boxes")
if self._shp is not None:
old_shp = tf.cast(tf.shape(image)[:2], dtype=tf.float32)
new_shp = tf.cast(self._shp, dtype=tf.float32)
image = tf.image.resize(image,
self._shp,
self._method,
antialias=True)
# resize bounding box
y, x, h, w = boxes[:, 0], boxes[:, 1], boxes[:, 2], boxes[:, 3]
y = new_shp[0] * y / old_shp[0]
x = new_shp[1] * x / old_shp[1]
h = new_shp[0] * h / old_shp[0]
w = new_shp[1] * w / old_shp[1]
coordinates = tf.stack([y, x, h, w], axis=1)
else:
coordinates = boxes[:, :4]
# one-hot classes
classes = boxes[:, 4]
if self._one_hot_class:
classes = tf.cast(classes, dtype=tf.int32)
classes = tf.one_hot(classes, depth=self.nClass)
else:
classes = classes[..., tf.newaxis]
boxes = tf.concat([coordinates, classes], axis=1)
example["image"] = image
example["boxes"] = boxes
return example
